Issues / #915

#915 Windows AMD report: RX 6800M (gfx1031, laptop) works with a self-built HIP engine — please add 0x73DF / gfx1031 to the Windows path

open · @Ultmate-hub · 1 Kommentare · Auf GitHub

BenchmarksSetup & installServer & APIAMD / HIPNVIDIA / CUDAModels & quantsWindows

Beschreibung

Strata 0.1.39 runs on a laptop **RX 6800M (Navi 22, gfx1031)** under Windows 11, using an engine I built with `tools\hip\build_windows.bat` for gfx1031. Setup needed one change: it doesn't recognise the card's PCI ID. Details below in case you want to add gfx1031 to the Windows build and release zip.

### PC
- ASUS ROG Strix G15 (G513QY), Windows 11 Home 23H2 (build 22631)
- GPU: AMD Radeon RX 6800M, 12 GB (PCI `VEN_1002&DEV_73DF`), plus the Ryzen's integrated Radeon (`DEV_1638`, correctly listed as not supported)
- AMD driver: 32.0.21045.5002 (8/17/2026), Adrenalin 26.8.1
- CPU: Ryzen 9 5980HX (AVX2), RAM: 64 GB
- Model: Qwen3.8-Flash-Next IQ2_XS, 32K context, 8-bit KV cache

### What blocked the normal install
- `START-HERE.bat --check --backend hip` reported `AMD Radeon RX 6800M, 12 GB VRAM - not supported ... unknown (PCI 73DF)`. `_WIN_AMD_DID` in `setup.py` has no 0x73DF, and the name regex for gfx1030 deliberately excludes `6800M`.
- The release `strata-windows-x64-hip.zip` is built for gfx1100/1101/1102/1200/1201/1030, so it has no gfx1031 code either.

### What I did
1. Added the PCI ID to `setup.py`:
   ```python
   0x73BF: "gfx1030", 0x73AF: "gfx1030", 0x73A5: "gfx1030",            # RX 6800 / 6800 XT / 6900 XT / 6950 XT
   0x73DF: "gfx1031"}                                                  # RX 6700 XT / 6750 XT, RX 6800M / 6850M XT
   ```
   (`AMD_ARCHS` already includes gfx1031.)
2. Built the engine with `STRATA_HIP_ARCHS=gfx1031`. TheRock publishes `rocm-sdk-device-gfx1031` 10.2.0a20260930 as a `win_amd64` wheel, so the pinned ROCm installed fine. The build itself had no errors (warnings only).
   - No git on this PC: I downloaded llama.cpp at `3cf03257` as a zip and passed `STRATA_GGML_DIR`.
   - The VS Build Tools CMake component wasn't installed, so I used `cmake` + `ninja` from pip in `.rocm-win`.
3. Installed it with `START-HERE.bat --yes --backend hip --prebuilt <dist folder> --family qwen --model IQ2_XS`.

Minor: on the first run, `package_windows.py` failed while zipping with `FileNotFoundError` for a staged `rocblas\library\TensileLibrary_..._fallback_gfx1031.dat`. Most of the staged rocBLAS files had disappeared from the stage folder. Rerunning only the packaging step worked (142 files, 119 MiB zip). I couldn't find a cause, and I can't reproduce it.

### `engine\strata-device.exe --list-devices`
```
device 0: AMD Radeon RX 6800M
  arch gfx1031, 12.0 GiB, wave32
```

### `engine\strata-device.exe --selftest` (run while the server was up)
```
device 0: AMD Radeon RX 6800M
  HIP arch            gfx1031 wave32 (compiled for gfx1031)
  multiprocessors     20
  VRAM total / free   11.984 GiB (12868124672 B) / 11.068 GiB (11884216320 B)
  driver / runtime    71726391 / 71726391

memory plan  --max-context 20480
  KV + indexer keys   0.267387 GB
  expert cache        5.675613 GB  (4105 slots of 1,382,400 B)
  VRAM pool           5.943000 GB
  plan + KV vs free   FITS

selftest: over-allocation refused as required
strata-device selftest OK
strata: Windows budgets 11474 of this card's 12272 MiB for this process
```

### End of `strata-iq2_xs.log` (startup)
```
strata generate: GPU 0: AMD Radeon RX 6800M (gfx1031)
strata generate: expert arena: locked 2951 MiB via working-set minimum + VirtualLock; cudaHostRegister of the whole arena FAILED (invalid argument); 44 slices pinned (30 GiB); large pages refused (VirtualAlloc error 1314); using 4 KB pages
strata generate: loaded 33.02 GiB at 1.27 GiB/s
strata generate: expert cache 4166 slots, 5.59 GiB of VRAM
strata generate: pre-filled 4166 of 4166 slots from the profile; slot 0 verified
strata generate: session is up (engine 0.1.39)
strata serve: the prompt path borrows 3123 CUDA0 cache slots (4.18 GiB)
strata serve: 462 MiB of VRAM free with everything loaded
```

### Speed (first answers, `reasoning_effort: none`)
```
prompt 25 tokens ... read in 1458 ms (17.1 tok/s), 8 generated in 639 ms (12.5 tok/s), drafts accepted 6 of 6
prompt 30 tokens ... read in 1454 ms (20.6 tok/s), 106 generated in 6332 ms (16.7 tok/s), drafts accepted 79 of 84
prompt 57 tokens ... read in 2615 ms (21.8 tok/s), 367 generated in 20387 ms (18.0 tok/s), drafts accepted 239 of 332
decode expert cache hit rate: 67.7%; 3.9% of routed experts read over PCIe
```
The answers were correct: a factual question, and a short Python `is_prime` function that is right.

### Side note on the driver
The PC first had a May 2023 driver (31.0.14057.5006). `strata-device --list-devices` found the card once on it. Later in the session the RX 6800M dropped into Device Manager code 31 (`CM_PROB_FAILED_ADD`), even after a reboot. I can't say whether Strata had anything to do with it. Updating to 32.0.21045.5002 fixed it, and everything above ran on the new driver.

### Request
Could you add gfx1031 to the Windows HIP release zip (`STRATA_HIP_ARCHS`) and `0x73DF` to `_WIN_AMD_DID`? Other Navi 22 IDs may be worth adding too. Then RX 6700 XT / 6750 XT / 6800M owners on Windows wouldn't need to build it themselves. I'm happy to test a build.

Mehr auf der Site

Links zu Install, Modellen, Releases.